caveman, claude code skill that cuts output tokens by 65%
About Caveman
Open-source skill that makes AI coding agents answer in terse caveman speak, cutting output tokens by 65% on average (22-87% by task) while keeping code, commands, and error strings byte-for-byte exact. Six intensity levels from lite to classical-Chinese wenyan. Works across Claude Code, Cursor, Windsurf, Cline, Gemini CLI, Copilot, and 30+ agents. 86.9k GitHub stars, MIT license, no telemetry. Note: compresses output tokens only; the skill adds roughly 1-1.5k input tokens per turn, so whole-session savings run smaller than the headline.
Key Capabilities
65% average output-token reduction, measured 22-87% across benchmark tasks
Code, CLI commands, API names, and error strings preserved byte-for-byte
Six intensity levels: lite, full, ultra, plus three classical-Chinese wenyan modes
One-command install for 30+ agents; on Claude Code it is a single SKILL.md file
Auto-Clarity rules drop compression for security warnings and destructive actions
No telemetry, no accounts, zero network calls after install (MIT, 86.9k stars)
Standout Features
65% Output Compression
Terse caveman speak drops articles, filler, and pleasantries. Repo benchmarks show 294 vs 1,214 output tokens on the same tasks.
Technical Substance Untouched
Code blocks, CLI commands, API names, and exact error strings stay verbatim. Only prose gets compressed.
Six Intensity Levels
From lite (tight but professional) through ultra (one word when one word enough) to wenyan classical-Chinese modes at 80-90% character reduction.
Auto-Clarity Safety Valve
Automatically reverts to full prose for security warnings, irreversible-action confirmations, and sequences where compression risks misreads.
Perfect for
Compare Caveman
Community Buzz
Related Articles
ready to try caveman?
join thousands of developers already using Caveman.
try caveman now